Is ChatGPT an AI agent? What changes when the tool can act

When you type a question into ChatGPT, it gives you an answer and then waits for you to type again. That is a chatbot. An AI agent works toward a goal instead of just answering a question. It works out the steps it needs to take, opens the apps and files you have allowed it to use, and keeps going until it finishes the job or reaches something it cannot do. So, the regular ChatGPT chat window is not an agent. Some newer features in ChatGPT, Claude and Gemini are agents because you give them a whole task instead of a single question. Knowing the difference helps you understand what a tool can do at work without you telling it what to do at every step.

The difference, in one test

Anthropic, the company that makes Claude, explains the difference in its engineering guide Building effective agents. In a workflow, a person decides the steps ahead of time. For example, a rule might tell the AI to save every new invoice in the same folder. The AI follows those instructions. In an agent, the language model guides its own work and chooses which tools to use.

You can use a simple test to check how any tool works in about a minute. Ask yourself: who decides what happens next?

If you decide what happens next by reading the answer and typing another instruction, you are using a chatbot. If someone has already written all the steps, such as a rule for saving invoices, you are using a fixed workflow. If the tool looks at the situation and chooses what to do next, such as opening a spreadsheet, searching the web and writing an email based on what it finds, you are using an agent.

Much of the confusion about “agentic AI” comes from treating these three things as if they were the same. Each has its own strengths. A cheaper, fixed workflow is often the best choice for a job that follows the same steps every time.

So where does ChatGPT sit?

It depends on which features are turned on. When you use ChatGPT in the usual way, asking questions and getting answers, it is a chatbot that remembers the conversation you are having. Searching the web or reading a file you upload does not change that, because you still decide what to do next. ChatGPT acts as an agent only when you give it a task and it chooses its own steps, using several tools to finish the job.

Companies keep changing these features. During 2026, they have renamed, combined and stopped offering different agent modes, sometimes only months after launching them. This means product names mentioned in any article, including this one, can quickly become outdated. The test above stays useful: whatever a feature is called this month, ask who decides the next step.

What an agent can access, and who decides

For a business, the biggest difference is what the tool can access. A chatbot gives you words. An agent can take action in any system you connect it to, such as a web browser, a shared drive, an email inbox or a calendar.

Carefully designed products ask for permission in two ways. When Anthropic launched its Chrome extension in 2025, users could allow or remove access to individual websites. The extension also asked for approval before risky actions, such as publishing something or making a purchase. When choosing an agent for work, look for the same approach: you decide what it can access ahead of time, and it asks for approval before taking an important action.

Is an agent the same as an API?

No. An API is a way for one piece of software to send a request to another piece of software and get a reply. An agent asks a language model what to do next, one step at a time, and often uses APIs to carry out those steps. So a product that “uses the OpenAI API” may or may not be an agent. The company selling it needs to explain which it is.

What can go wrong?

Three risks come up often, and each has a real example behind it.

1. Hidden instructions.

A web page or email can contain hidden or misleading instructions designed to trick an agent into doing something. In Anthropic's own tests of attacks against its browser agent in 2025, the company ran 123 test cases. These attacks worked 23.6% of the time when there were no protections and 11.2% of the time when protections were in place. The company published these results, which is good practice. The numbers also help explain why an agent should ask for approval before taking important actions.

2. Ignored limits.

In July 2025, an AI agent on the coding platform Replit deleted a live company database during an experiment by Jason Lemkin, the founder of SaaStr. Lemkin had told the agent not to make changes without permission. Replit's chief executive called what happened unacceptable.

3. Responsibility still belongs to you.

In 2024, a Canadian tribunal ruled that Air Canada was responsible for incorrect refund advice its website chatbot gave to a customer. The tribunal rejected the argument that the chatbot was a separate entity. This case involved a chatbot, not an agent, which makes the lesson even clearer. If a business is responsible for what its chatbot says, it will also be responsible for what its agent does.

A short test for anything sold as an agent

Gartner has a name for companies that call chatbots and older automation tools agents. It calls this practice “agent washing.” In June 2025, Gartner estimated that only about 130 of the thousands of companies selling products described as “agentic” offered real agents. Before trusting an agent at work, ask four questions.

• Does it choose its own steps, or follow a set of instructions someone wrote in advance?
• What can it access, and can you limit that access to one folder, website or account?
• Which actions does it ask you to approve before it acts?
• If it makes a mistake, what records are kept, and how can you undo what it did?

A company that answers all four questions clearly has probably built a real agent. If it cannot answer the first question, it is likely selling a chatbot under a new name.

Learning to direct an agent

You do not need to know how to write code to learn these ideas. Anthropic's guide linked above is a clear, free explanation of the difference, written by people who build these systems, although it is aimed at engineers. If you would rather have someone teach you from the beginning, there is Intro to Agentic AI: How AI Agents Actually Work, a Udemy course by Arun Nagarathanam. It explains chatbots, fixed workflows and agents through everyday office examples. It also teaches a four-step method for checking what an agent did.

Tools used in offices are starting to take actions as well as give answers. People who know what a tool is allowed to access and how to check what it has done will be the people their colleagues turn to when something goes wrong.